100 Hours Testing Claude Code vs ChatGPT Codex — Honest Results

Key Takeaways

• Claude Code and Codex share more than they differ: both support skills (markdown files), MCP, CLI, desktop apps, VS Code extensions, hooks, sub-agents, and cloud delegation.
• Claude Code wins on customization depth: 30 hook events vs Codex's 6, auto-spawning sub-agents, /ultraplan, /ultrareview, /loop, Channels MCP server, and the Agent SDK.
• Codex wins on unified shipping: built-in git work trees for parallel tasks, in-app browser with visual commenting, and sharper computer-use QA flows — all in one window.
• Claude Code feels "more creative and pushes back when you're going wrong"; Codex feels "sharper at following instructions and reviewing code for bugs."
• Codex is included in every ChatGPT plan (including free); Claude Code requires a paid Anthropic subscription — a meaningful difference for teams already on ChatGPT.

Where Claude Code Leads

The hook system is the clearest differentiator. Claude Code offers 30 hook events — automated triggers that fire on prompt submission, tool calls, session starts, task creation, and more. Codex currently has about six. If you want to wire automated behavior into every stage of an agent's workflow, Claude Code gives roughly 5x the granularity. Beyond hooks, Claude Code's auto-spawning sub-agents let it spin up a planner, explorer, and reviewer on its own for complex tasks, without you having to ask. Codex's docs explicitly say sub-agents require explicit invocation.

The /ultraplan and /ultrareview slash commands — both in research preview — let you ship the planning and review phases to cloud Claude Code sessions and review them in a browser with inline comments. /loop puts Claude in maintenance mode to handle unfinished tasks and PR comments on a recurring schedule. These have no Codex equivalents today.

Where Codex Leads

Codex is built around git work trees from the ground up — every task thread gets its own working copy of the project, so parallel work never collides. Combined with review, stage, commit, and push all inside the same desktop app, it functions as a complete shipping pipeline out of the box. The in-app browser lets you visually comment on the page your agent just built without switching to Chrome. Its computer-use QA flow (tell Codex to QA your app, it clicks around and logs bugs with severity ratings) is described as "really sharp."

Commands & Code Mentioned

/ultraplan
/ultrareview
/loop
channels (MCP server)
claude-agent-sdk

Related on OpenClawDatabase

• OpenClaw (Claude Code) Hub — setup, configuration, and skills guides
• Compare agent platforms — side-by-side comparison tool
• ChatGPT / Codex guide — Codex setup and overview